I built a YouTube transcript exporter, and one design question kept coming back: when the timedtext endpoint won't cooperate, is the on-page transcript panel a real fallback — or does it only hold the rows currently on screen?

On 2026-11-15 I measured that YouTube's timedtext endpoint returns empty bodies from datacenter IPs (curl: HTTP 200 with 0 bytes, then 429; real browser: json3 through srv3 all empty, no challenge page). Real users on residential IPs should be fine, but my test environment could not verify that path. So the panel became the question that mattered. I originally wrote that the panel "only gives the currently rendered rows, not the full transcript." That was an untested assumption. I tested it properly this week, at three scales.

How I measured it

Headless Chrome 154 with a local CONNECT relay, loading the real youtube.com watch page. For each video I opened the transcript panel, counted the rows, then scrolled the panel container to the bottom twice and counted again. I also ran my extension's own YTE_GET_TRANSCRIPT (source: panel) before and after scrolling, so the panel DOM and the extension's reading of it were checked against each other.

Results

Video Length Rows before scrolling Rows after 2 scrolls Cues (extension) Characters
Rick Astley — Never Gonna Give You Up (dQw4w9WgXcQ) 3:33 24 24 24 2,066
TED — Do schools kill creativity? (iG9CE55wbtY) 20:03 163 163 163 17,412
Randy Pausch — The Last Lecture (ji5_MqicxSo) 76:27 713 713 713 65,510

First cue to last cue covered the whole video in each case (0:01–3:23, 0:27–19:19, 0:17–1:15:26). The panel container's scrollHeight was far larger than its clientHeight (e.g. 39,870 vs 299 on the 76-minute video), so there was real scrolling to do — the row count simply did not change. The extension's transcript was identical before and after scrolling in all three runs.

What changed in my design

The panel is now the default full-transcript source (verified up to 76 minutes / 713 cues). timedtext is demoted to a backup for when the panel is unavailable. The panel DOM itself also changed recently: the old ytd-transcript-segment-renderer selector returns zero rows now; the current rendering uses <transcript-segment-view-model> with timestamps in .ytwTranscriptSegmentViewModelTimestamp and text in span[role="text"]. My reader tries the new DOM first and falls back to the old one.

Honest caveats

  • I have not tested 2-hour-plus videos or 1,000+ cue transcripts. If a virtualization threshold exists, it can only be at a larger scale than 713 cues.
  • My DOM probe truncated each row to 80 characters, so its character totals (1,776 / 11,471 / 49,505) are lower than the extension's (2,066 / 17,412 / 65,510). That is a probe-method difference, not a data difference.
  • The 2026-11-15 claim that "the panel only gives the currently rendered rows" is overturned at all three scales tested. It was wrong because I had not scrolled before writing it.

Status

YouTube Transcript Export v0.1.0 is in review on Microsoft Edge Add-ons and cannot be installed yet. The Pro tier (planned ¥19.90 one-time, Afdian activation code) is not on sale. Privacy page: https://alapha888.github.io/yt-transcript-export-privacy/